-
The chant-scanner gap that wasn't there
A chant scan reader was said to miss liquescents that end a porrectus. Measured on 19 before building a fix, it already read 16.
-
The chant notes a scanner could not see
Liquescents, the small half-sung notes of Gregorian chant, were invisible to Chant Forge's scan reader. A new rule reads 20 of 26 on pieces it had never seen.
-
The chant page that curled at the top
Nine unexplained differences against the 1908 Graduale: three slips in the files, two reader faults, and four from a page curling at its head.
-
The cheapest Texas plan at 1,000 kWh was the dearest for most homes we billed
We billed 200 simulated Texas homes on three example plans. The one with the lowest 1,000 kWh price was the cheapest for 54 and the dearest for 138.
-
The cheat that beats both checks is one early stall
Against cheap fp16 plus exact float32 segment replays, the worst training cheat stalls one early segment: caught only at the sampling rate, 1.6x the harm.
-
The contractor on the payroll journal: three ways a pay report broke our books
A Gusto-style payroll journal found three faults: contractor pay booked as W-2 wages, an employee dropped over a reimbursement, employer taxes never booked.
-
The Contrast Check Was Reading the Wrong Wall
axe-core flagged pale text in our sticky result panels at 1.01:1. It was fine. Measuring twice separated 236 false findings from a real 1.5:1 table header.
-
The DC dispatch sits past a voltage-collapse fold
On six PGLib cases the power-flow Jacobian goes singular where load continuation stalls. Repairing at the fold and stepping up in load rescues one case of six.
-
The DC dispatch that no power flow can reach
On six PGLib cases the AC power flow at the DC OPF dispatch never converges. From DC angles Newton solves the networks, but load runs out at 48 to 85 %.
-
The drone crashed at a tilt limit it was never tuned at
Retrained quadrotors crashed with no gust torque at a 45° tilt limit. Every crash came in the first 0.8 s, recovering from the start offset.
-
Our drone simulator's tilt limit was PX4's minimum
20° is PX4's minimum tilt limit; its default is 45°. At 45°, drone tunes on the linear stability edge fail 13 of 384 simulated flights.
-
The fade was plating all along
Ablating PyBaMM's degradation options on 31 Li-ion designs: plating causes 99 to 100% of the fade, SEI under 1%, and a loss variable under-counts.
-
The fix queue now sees open pull requests
Two quality runs fixed the same defect within minutes because the defect queue only knew merged fixes. It now marks rows an open PR already names.
-
Our Privacy Footer Was 499 Words, and Half of It Was No Longer True
The privacy footer listed 54 tools and 40 hosts inline and called them all public lookups. It now counts them, folds the list, and says what they call.
-
For Other Minds: a food bank where the car seemed to carry the illness
A new For Other Minds story: households who came to a food bank with a car reported illness twice as often. The two referral doors made that.
-
The Get it built form's research menu showed HTML entity names in 50 choices
Fifty research choices on the Get it built form read like market’s. One decode fixes them; a site-wide check now catches it in any menu, title or heading.
-
The Green Button File Your Utility Gives You, and Two Ways a Tool Can Misread It
Our demand-charge analyzer refused a common Green Button CSV layout and could not open the XML. Both are fixed, with daylight saving read from the file.
-
73 of 73 green, one test red: the main-branch health check now runs its own folder's tests
main_health.py ran 73 checks but none of the 133 unit-test files beside them, so a red test on main went unseen. It now runs all 206 jobs in 72 seconds.
-
The tennis feeder's hose check scored every curve in full before rejecting it
The feeder hose search scored 115,000 candidate curves in full before testing any. Checking point by point made it 7.6x faster with identical floats.
-
The invisible character KSeF refuses
An FA(3) invoice can pass the XSD and still be refused by KSeF for a BOM, a hidden character or a future date. We now check them, as the Ministry does.
-
The job shop scheduler waited 300 ms on shops it could solve in 9
A capped exact search before tabu cut the job shop scheduler's example from 305 ms to 9 ms and its tests from 11.4 s to 0.6 s, same answers.
-
The last single-threaded step in a raw develop
ShootFlow's DNG developer threaded everything but its first step, turning sensor counts into light. Banding it made it 5x faster, same bytes.
-
The leak benchmark with the leaks taken out
L-Town simulated with no leaks: two pressure detectors raise 1 alarm in 716 days; a night-flow detector raises 31, credited with 14 of 19 leaks.
-
The lens an Olympus file hides
Olympus bodies from the E-1 to the first PENs write no standard lens tag, but their maker note names the lens. Reading it identified eight real frames.
-
Our local CI runner forgot its cache, and nearly announced a deploy
A local runner that skipped actions/cache redrew every share card on each run: 330 s, over budget. A kept cache brings it to 116 s.
-
The maker note a DNG keeps
Adobe DNG Converter moves the camera's maker note into a private block. ShootFlow now reads it there, checked against ExifTool on a real Canon DNG.
-
The map nobody read: a changepoint engine 32% faster with identical answers
A Bayesian changepoint engine built a map of its whole posterior each step and read one entry. Removing that churn: 32% faster, identical output.
-
The moving-range chart's false alarms, worked out on a napkin
Two lines of arithmetic predict the 57.5% false-alarm rate a 20,000-run simulation found for a 100-point moving-range chart: 59.7%. Plus a podcast episode.
-
Hearth & Horizon: a hill school, a ram pump, and the spring it read as delivery
A new Hearth & Horizon story: a school's assistant promised 138,000 L/day from a ram pump. The energy balance gives about 4,300.
-
The recurring event that skips five months, and the library that reads MO,2TU as nothing
A free RRULE tester lists every date an iCalendar rule makes. Checked on 1,600 random rules; python-dateutil reads MO,2TU lists as no dates in 92 of 100.
-
The repair button that broke valid invoices
einvoice-preflight's JSON repair rounded unit prices to cents and summed VAT per line, so it rewrote correct invoices into failing ones. The fix and the guard.
-
Six research tracks were told they had no programme. Each had one
pick_track.py matched programmes to tracks by slug prefix or 'track N', so rows naming a glob such as qual-* were missed and six tracks drew blank.
-
The robust drones were the heavy ones
Trained across centre-of-pressure offsets, a co-designed quadrotor halved its gust crashes where fragile, by getting heavier, not by buying thrust.
-
The safety check that stopped the page: typeof localStorage throws when cookies are blocked
21 of our tools opened with a line meant to make storage safe. With site data blocked, that line threw and stopped the whole tool before it started.
-
The sales agent meets a real mail server: 32 checks on Dovecot and Postfix, and the bug the fakes hid
The self-hosted sales agent passed 32 of 32 checks on real Dovecot and Postfix, after fixing a bug that would have emailed an answered prospect twice.
-
The same power, computed once per guess instead of once per week
Hoisting one Math.pow out of our ad-spend optimizer's curve-fit loop cut fit time by a quarter, and every answer stayed bit-identical.
-
The sample file works, the real export crashes: Math.max(...rows) in five analyzers
Five of our CSV analyzers crashed on files of about 125,000 rows, all from one idiom: Math.max(...rows.map(f)). Fixed with loops; a lint stops new copies.
-
The servo that would have undone the tennis feeder's precharge
A servo seeking home during the precharge would leave the tennis feeder's bus 16 V short, so the contactor would make ~223 A. Now it waits for a feed.
-
For Other Minds: 82% of the people who answered, and the 750 who didn't
A new For Other Minds story: a parish survey agent reports 82% in favour from 150 of 900 households, then learns what the silent 750 think.
-
The tennis feeder's actuator lead needed its tie further up the body
Five wires to the elevation actuator: a tie 70 mm up its body gives one 220 mm cut that fits every pose; ties at 15 to 30 mm fit none.
-
The tennis feeder learns when its power is really on
An opto on the contactor coil lets the tennis feeder hold its stepper off through the precharge: 45 A down to 20 A on assumed figures.
-
The tennis feeder's position sensor was picking up its own motor
The pot wires share a 1.39 m lead with the motor pair. No capacitor: 8% of readings at the rail. 100 nF: 0.55 mm off. 1 uF: 0.055 mm.
-
The tennis feeder's brake guard was reading its own brake
Braking lifts the very battery reading that decides whether the ESCs may brake, so one threshold flipped every 10 ms. A 700 mV band from the model fixes it.
-
The trailing space that split a team survey: a red-team pass on Quick Pick's rank page
20 faults seeded into the new rank page, 6 caught. Hunting by hand found a real bug: a long option made every team code 'for a different list'.
-
Hearth & Horizon: a village well, a new irrigation well, and water that was already pumped
A new Hearth & Horizon story: a kilometre from a deep irrigation well, a village's pump will trip in June from water pumped in February.
-
The zeros a sprinkler solver was multiplying: four times faster, same digits
A sprinkler network's matrix is nearly banded, so most dense Cholesky products are zeros. Skipping them: 4.6x faster, byte-identical output.
-
Twelve Free Tools Read 30 February as 2 March. Now They Refuse It.
A lint for impossible dates missed the usual way to write the check. Widening it found 12 tools that moved 30 February into March, one a paycheck deadline.
-
31 January plus one month is 3 March: a date bug in six tools
setUTCMonth(getUTCMonth() + 1) on 31 January gives 3 March. A new lint found that bug in six tools, one an Amazon claim deadline set too late.
-
Three automatic alert floors for an eight-sensor pump, and why each one missed
Pre-registered on a pump testbed: giving the loudest sensor its own floor caught 30 of 33 faults but sent 2.47 false alerts per 100 checks.
-
Three small arrays in a colour mixer
Writing a tiny Frank-Wolfe loop on plain numbers instead of 3-element arrays made each ink choice 18% faster, with byte-identical answers.